The last-mile gap is where enterprise AI projects die: the customer’s data is messy and undocumented, their systems are old, their security team has not approved anything, and nobody defined what “working” means precisely enough to test. None of these are model problems, which is why buying a better model does not close the gap.
Worked example: MIT’s NANDA study found roughly 95% of enterprise AI pilots produced no measurable P&L impact — and a large share of those systems worked, but could not demonstrate it because no baseline or metric was agreed in advance. Gotcha: the gap is mostly organisational, so the skills that close it — scoping, negotiated evals, stakeholder work — are the ones engineers are least likely to have practised.